Seeds — rolling the same dice twice
Every image generation starts from a field of random noise, and the seed is the number that picks which noise. Fix the seed and you fix the starting point: same seed + same prompt + same model version → the same render, near-pixel-identical, today and next month. Change the seed and you get a genuinely new draw of the same idea.
Run the editor. Two random generators seeded with 4200 produce identical sequences; the one seeded with 7 goes its own way. That's the whole mechanism — image models just use the sequence to fill a canvas with noise instead of a list with integers.
Why a working designer cares:
- Reproducibility is client insurance. "The client approved v3 of the hero — can we get it again with the tagline area clear?" If you logged the seed, yes, exactly. If you didn't, you're re-rolling and praying the new draw resembles the approved one. Log
(prompt, model version, seed)for anything a client says yes to. That triple is the approval. - Seeds isolate your experiments. Keep the seed fixed and change one prompt token — the difference between renders is caused by your edit, nothing else. It's the iteration loop from lesson 1 with the noise variable controlled. Change seed and prompt at once and you learn nothing, which is the slot-machine loop wearing a lab coat.
- Planned seeds structure a batch. For fifty variants you don't want one seed (fifty near-identical images) or random seeds (fifty strangers). You want a plan: base seed, incremented per variant —
4200, 4201, 4202…— so every asset is a distinct draw, every asset is reproducible, and the log reads like a contact sheet instead of a mystery.
Two honest caveats. First: seed behavior is per-model, and a seed only pins the model's starting noise — a big prompt change still lands somewhere new. Second: "same model" means same model version. Hosted services bump the default under you without asking, and even on the same version GPU arithmetic allows a whisper of nondeterminism — expect near-pixel-identical, not bit-identical. So the log records the version string, and "exactly" means "no client will ever see the difference." Seeds are the lock on the dice, not on the whole casino. That's why they're the fourth lock, not the only one.
The predict drill next: read a batch plan and say exactly which seed each deliverable gets.